skills/transformation/depth-estimation falls through to torch, a dependency the macOS deploy path deliberately excludes

[solderzzc]

skills/transformation/depth-estimation falls through to a dependency the macOS deploy path deliberately excludes

Same bug class as #208 (YOLO detection): the accelerated-backend fallback chain has no path for "installed correctly, but the accelerated backend is unavailable here" — it eventually falls through to a dependency (torch) that the lightweight macOS deploy.sh path deliberately does not install.

Repro

Clean venv matching deploy.sh's macOS branch: coremltools 9.0, huggingface_hub, numpy 2.5.3, opencv-python-headless, Pillow, matplotlib — no onnxruntime/torch/depth-anything-v2.

Forced a CoreML load failure without touching the repo (pointed MODELS_DIR at a throwaway temp dir, no-op'd the HF download). The full fallback chain fires as designed, then dies unguarded:

CoreML load failed → falling back to PyTorch
ONNX Runtime load failed (No module named 'onnxruntime') → trying TensorRT
TensorRT unavailable (No module named 'torch') → falling back to PyTorch
  File ".../depth-estimation/scripts/transform.py", line 464, in _load_pytorch
      import torch
  ModuleNotFoundError: No module named 'torch'

The final fallback call (_load_pytorch) is unguarded — no try/except — so this ModuleNotFoundError propagates straight to the caller instead of a clean "no usable backend" error.

Secondary note

deploy.sh's macOS branch has no numpy<2.0 cap, unlike requirements_mps.txt for the YOLO skill (which pins numpy>=1.24.0,<2.0.0). Worth checking whether that's intentional or a consistency gap between the two skills' deploy paths.

Same fix shape as #208 likely applies here too: don't let the fallback chain reach torch on a deploy path that never installs it — fail with a clear "no accelerated backend available" message instead, or check framework_ok-equivalent before attempting the torch import.

🤖 Generated with Claude Code